memory issue in RPA calculation

Run-time issues concerning Yambo that are not covered in the above forums.

Moderators: Daniele Varsano, andrea marini, Conor Hogan, myrta gruning

Re: memory issue in RPA calculation

Postby eariel99 » Mon May 14, 2012 11:12 am

>Fine. But be careful with the PBE. Check when running p2y which kind of XC functional is using >Yambo. Most probably it using some LDA-like.

I had noted that yambo reports that is it using LDA. As for RPA dielectric function all that is needed is to integrate the wavefunctions, I thought it does not matters what functional was used to generate them.

In the case of GW eigenvalues, I thought that yambo calculates the LDA igenvalues and the GW corrections using the slightly different PBE wavefunctions. As I have never met any great difference between PBE and LDA I do not worry too much. Should I ?
Thanks, Eduardo
Eduardo Menendez Proupin
Departamento de Química Fisica Aplicada
Facultad de Ciencias
Universidad Autónoma de Madrid
28049 Madrid, Spain
On leave from: Departamento de Fisica, Facultad de Ciencias, Universidad de Chile URL: www.gnm.cl/emenendez
eariel99
 
Posts: 10
Joined: Tue Apr 24, 2012 10:35 pm

Re: memory issue in RPA calculation

Postby eariel99 » Mon May 14, 2012 11:17 am

Aaahh!! OK. The velocity commutator is different for LDA and PBE, and this may be the reason to have bad results for CdTe. My test with Si were with LDA all the time.
Eduardo
Eduardo Menendez Proupin
Departamento de Química Fisica Aplicada
Facultad de Ciencias
Universidad Autónoma de Madrid
28049 Madrid, Spain
On leave from: Departamento de Fisica, Facultad de Ciencias, Universidad de Chile URL: www.gnm.cl/emenendez
eariel99
 
Posts: 10
Joined: Tue Apr 24, 2012 10:35 pm

Re: memory issue in RPA calculation

Postby andrea marini » Mon May 14, 2012 11:21 am

eariel99 wrote: I had noted that yambo reports that is it using LDA. As for RPA dielectric function all that is needed is to integrate the wavefunctions, I thought it does not matters what functional was used to generate them.


It depends on how different is the wavefunction. If the LDA and PBE functions are very different you should see a difference in both the RPA and the GW corrections.

eariel99 wrote: As I have never met any great difference between PBE and LDA I do not worry too much. Should I ?


How did you judge that differences between LDA and PBE are small ? Can you check the matrix elements of the xc potential, for example ?

eariel99 wrote: The velocity commutator is different for LDA and PBE, and this may be the reason to have bad results for CdTe. My test with Si were with LDA all the time.


Hmmm. Yoi are right. The PBE functional depends also on the gradient. Therefore the expression for the commutator will change. GOOD OBSERVATION!

Ok. Step by step ....

Can you first do a pure LDA calculation and check that the effect of the commutator is small ? And, regarding doing GW on top of PBE if you are interested in doing it properly drop me an E-mail.

Andrea
Andrea MARINI
Istituto di Struttura della Materia, CNR, (Italy)
User avatar
andrea marini
 
Posts: 447
Joined: Mon Mar 16, 2009 4:27 pm

Re: memory issue in RPA calculation

Postby eariel99 » Mon May 14, 2012 2:27 pm

Hi Andrea,
Thanks again for your answer. Let me answer and ask more.

How did you judge that differences between LDA and PBE are small ? Can you check the matrix elements of the xc potential, for example ?

I have not, in this specific case. I only know iron as example that LDA predicts a wrong crystalline structure, and in my work I have not found significant differences in band diagrams and other magnitudes when I have used both LDA and PBE. I have never evaluated numerical differences between wavefunnction. Of course, it is not consistent to mix the PBE wave functions with the LDA calculation.

Can you first do a pure LDA calculation and check that the effect of the commutator is small ?

Yes. this is the calculation for silicon. For PBE with nonlocal contributions the order of magnitude is terrible, although the shape is OK.
compara-ld-pbe.jpg


For CdTe, the LDA calculations are here. The NL calculation increases somewhat the values, the shape is not bad. With PBE the shape was absolutely wrong, as seen in a former post.
compara-cdte-lda.jpg




And, regarding doing GW on top of PBE if you are interested in doing it properly drop me an E-mail.

I think it is safer to swith to LDA. I have too many atoms to calculate and the complications with PBE are too much for me. In any case, I have to generate new pseudpotentials for Cd that include 4s and 4p in the valence, and the task for LDA and PBE is the same. For the moment I just want to proof that I can obtain the quasiparticle gap in the known case of CdTe.

Please, let me ask a few more questions about the GW calculations, even if its beyond the title of this post. I am following again the tutorial of Basic concepts of the GW approximation (http://www.yambo-code.org/tutorials/GW/index.php). Three weeks ago in Lausanne I was primarily worried about compiling yambo with quantum espresso and I miss this part.

In the section of COHSEX, one reads that firsts it is needed the static screened interaction ( I guess, W(r,r',E=0 ) . For that one needs yambo -b and quasiparticle energies with yambo -b -g n. I am confused because it appears in a paragraph and later on, the COHSEX calculation is configured with yambo -g n -p c (without -b ). Then I have a few doubts

1) Does yambo -b -g n makes a SEX calculation, i.e., without the CO part?
2) Is it the same to perform
yambo -b;yambo;yambo -g n;yambo
as
yambo -b -g n;yambo
?

3) Do -b and -x act like adding ingredients to the calculations? If I do a COHSEX calculation without
calculating first the Hartree-Fock self-energy and the static screeing, what happens? Are them calculated on the fly ?
Note: I see that "yambo -p c -g n -F test-cohsex" and "yambo -x -b -p c -g n -F test-cohsex-x-b" produce the same input file but the order of some lines is different, I guess the order does not matter. Also,
"yambo -p p -g n -F test-ppa" and "yambo -x -b -p p -g n -F test-ppa-x-b" produce the same inputs.
but
"yambo -b -o b -y h" (from the CPC yambo article) produce a different input file than "yambo -o b -y h".
So, is there a general rule or a flexibility for BSE?

4) What is the interest in doing the COHSEX with empty bands? Is it to fix the parameter GbndRnge for later use in PPA or real axis calculation?

5) Hi Andrea,
Thanks again for your answer. Let me answer and ask more.

How did you judge that differences between LDA and PBE are small ? Can you check the matrix elements of the xc potential, for example ?

I have not, in this specific case. I only know iron as example that LDA predicts a wrong crystalline structure, and in my work I have not found significant differences in band diagrams and other magnitudes when I have used both LDA and PBE. I have never evaluated numerical differences between wavefunnction. Of course, it is not consistent to mix the PBE wave functions with the LDA calculation.

Can you first do a pure LDA calculation and check that the effect of the commutator is small ?

Yes. this is the calculation for silicon. For PBE with nonlocal contributions the order of magnitude is terrible, although the shape is OK.
compara-ld-pbe.jpg


For CdTe, the LDA calculations are here. The NL calculation increases somewhat the values, the shape is not bad. With PBE the shape was absolutely wrong, as seen in a former post.
compara-cdte-lda.jpg




And, regarding doing GW on top of PBE if you are interested in doing it properly drop me an E-mail.

I think it is safer to swith to LDA. I have too many atoms to calculate and the complications with PBE are too much for me. In any case, I have to generate new pseudpotentials for Cd that include 4s and 4p in the valence, and the task for LDA and PBE is the same. For the moment I just want to proof that I can to obtain the quasiparticle gap in the known case of CdTe.

Please, let me ask a few more questions about the GW calculations, even if its beyond the title of this post. I am following again the tutorial of Basic concepts of the GW approximation (http://www.yambo-code.org/tutorials/GW/index.php). Three weeks ago in Lausanne I was primarily worried about compiling yambo with quantum espresso and I miss this part.

In the section of COHSEX, one reads that firsts it is needed the static screened interaction ( I guess, W(r,r',E=0 ) . For that one needs yambo -b and quasiparticle energies with yambo -b -g n. I am confused because it appears in a paragraph and later on, the COHSEX calculation is configured with yambo -g n -p c (without -b ). Then I have a few doubts

1) Does yambo -b -g n makes a SEX calculation, i.e., without the CO part?
2) Is it the same to perform
yambo -b;yambo;yambo -g n;yambo
as
yambo -b -g n;yambo
?

3) Do -b and -x act like adding ingredients to the calculations? If I do a COHSEX calculation without
calculating first the Hartree-Fock self-energy and the static screeing, what happens? Are them calculated on the fly ?
Note: I see that "yambo -p c -g n -F test-cohsex" and "yambo -x -b -p c -g n -F test-cohsex-x-b" produce the same input file but the order of some lines is different, I guess the order does not matter. Also,
"yambo -p p -g n -F test-ppa" and "yambo -x -b -p p -g n -F test-ppa-x-b" produce the same inputs.
but
"yambo -b -o b -y h" (from the CPC yambo article) produce a different input file than "yambo -o b -y h".
So, is there a general rule or a flexibility for BSE?

4) What is the interest in doing the COHSEX with empty bands? Is it to fix the parameter GbndRnge for later use in PPA or real axis calculation?

5) And turning back to the Im(eps) calculation. Can I obtain the GW dielectric function using the G0W0 energies, and not with the scissor operator and valence/conduction stretching via %XfnQP_E? How ?

Best regards
Eduardo
You do not have the required permissions to view the files attached to this post.
Eduardo Menendez Proupin
Departamento de Química Fisica Aplicada
Facultad de Ciencias
Universidad Autónoma de Madrid
28049 Madrid, Spain
On leave from: Departamento de Fisica, Facultad de Ciencias, Universidad de Chile URL: www.gnm.cl/emenendez
eariel99
 
Posts: 10
Joined: Tue Apr 24, 2012 10:35 pm

Re: memory issue in RPA calculation

Postby eariel99 » Mon May 14, 2012 2:31 pm

opps, I accidentally pasted the clipboard on 5 and my post was almost doubles. The final question was just

5) And turning back to the Im(eps) calculation. Can I obtain the GW dielectric function using the G0W0 energies, and not with the scissor operator and valence/conduction stretching via %XfnQP_E? How ?
Eduardo Menendez Proupin
Departamento de Química Fisica Aplicada
Facultad de Ciencias
Universidad Autónoma de Madrid
28049 Madrid, Spain
On leave from: Departamento de Fisica, Facultad de Ciencias, Universidad de Chile URL: www.gnm.cl/emenendez
eariel99
 
Posts: 10
Joined: Tue Apr 24, 2012 10:35 pm

Re: memory issue in RPA calculation

Postby andrea marini » Wed May 16, 2012 8:30 am

eariel99 wrote: I have not, in this specific case. I only know iron as example that LDA predicts a wrong crystalline structure, and in my work I have not found significant differences in band diagrams and other magnitudes when I have used both LDA and PBE. I have never evaluated numerical differences between wavefunnction. Of course, it is not consistent to mix the PBE wave functions with the LDA calculation.


Ok. The check of the matrix elements of the PBE vs LDA is a test to do... but if you do not see any differnce in the energies maybe it is not that important. And you are absolutely right regarding mixing LDA and PBE!

eariel99 wrote:Yes. this is the calculation for silicon. For PBE with nonlocal contributions the order of magnitude is terrible, although the shape is OK.


GREAT TEST! Thanks! I have sent your post to the other developers. We will look into the problem shortly. PBE is now supported in the devel version of the GPL source that we will release in a matter of week. But thanks to your post we will, in the meantime, switch off the calculation of the non-local commutator.

eariel99 wrote:I think it is safer to swith to LDA. I have too many atoms to calculate and the complications with PBE are too much for me. In any case, I have to generate new pseudpotentials for Cd that include 4s and 4p in the valence, and the task for LDA and PBE is the same. For the moment I just want to proof that I can obtain the quasiparticle gap in the known case of CdTe.


That's fine. In any case in matter of week we will release the new GPL source that supports GW on-top of GGA-like xc-functionals.

eariel99 wrote:In the section of COHSEX, one reads that firsts it is needed the static screened interaction ( I guess, W(r,r',E=0 ) . For that one needs yambo -b and quasiparticle energies with yambo -b -g n. I am confused because it appears in a paragraph and later on, the COHSEX calculation is configured with yambo -g n -p c (without -b ). Then I have a few doubts

1) Does yambo -b -g n makes a SEX calculation, i.e., without the CO part?


No. Carefull. Each option corresponds to a piece of the run. I mean -b means "Yambo calculates W(r,r',E=0)". -g n means I do GW using the Newton solver. -p c means the kind of GW I do is COHSEX. -p p the kind of GW is Plasmon-pole.

Therefore yambo -g n -p c does COHSEX, or COH+SEX.

eariel99 wrote:2) Is it the same to perform
yambo -b;yambo;yambo -g n;yambo
as
yambo -b -g n;yambo ?


yes. Try yourself it this is the case.

eariel99 wrote:3) Do -b and -x act like adding ingredients to the calculations? If I do a COHSEX calculation without
calculating first the Hartree-Fock self-energy and the static screeing, what happens? Are them calculated on the fly ?
Note: I see that "yambo -p c -g n -F test-cohsex" and "yambo -x -b -p c -g n -F test-cohsex-x-b" produce the same input file but the order of some lines is different, I guess the order does not matter. Also,
"yambo -p p -g n -F test-ppa" and "yambo -x -b -p p -g n -F test-ppa-x-b" produce the same inputs.
but
"yambo -b -o b -y h" (from the CPC yambo article) produce a different input file than "yambo -o b -y h".
So, is there a general rule or a flexibility for BSE?


WoW. Too many questions :) Ok, step by step

1. The order of the variables in the input file does not matter
2. each option is a piece. -x does HF. -b calculates the statically screened interaction.
3. Sometimes Yambo selects automatically some logicals. For example if you want to do COHSEX you do need HF, therefore Yambo switches on automatically -x. These bindings are hard coded inside the code.
4. Similary you cannot do BSE without the statically screened interaction. Therefore when you do -o b yambo switches on also the -b option.

Is it more clear now?

eariel99 wrote: 4) What is the interest in doing the COHSEX with empty bands? Is it to fix the parameter GbndRnge for later use in PPA or real axis calculation?


This is an idea that must be checkd on the ground. Also because there is no actual energy dependence in COHSEX. But you can try. Run the COHSEX using empties and compare the convergence as a function of the empty bands with the PPA. And let us know ;)

eariel99 wrote:5) And turning back to the Im(eps) calculation. Can I obtain the GW dielectric function using the G0W0 energies, and not with the scissor operator and valence/conduction stretching via %XfnQP_E? How ?


If you did a GoWo calculation that you should have somewhere a ndb.QP databases. Let's assume it is in the TEST folder. Then to load the GoWo energies insted of doing a scissor f by using the -V qp verbosity and the QPdb option. In this case you should use

Code: Select all
"E < TEST/ndb.QP"


Andrea
Andrea MARINI
Istituto di Struttura della Materia, CNR, (Italy)
User avatar
andrea marini
 
Posts: 447
Joined: Mon Mar 16, 2009 4:27 pm

Re: memory issue in RPA calculation

Postby eariel99 » Wed May 16, 2012 4:06 pm

I am sorry, I did not understand what calculation is done by
yambo -b -g n;yambo

I got confused reading
In order to calculate the COHSEX self-energy with yambo you need first the static screened interaction yambo -b. As for the V(r) case, also W(r,r',E=0) will depend from the number of G-vectors, and the k-grid, but this time also the number of conduction bands XsBndsRn will enter in the screening calculation through the polarization function, see The Interacting response function: Many-Body and TDDFT section. Notice however that there are not any additional dependence of the self-energy operator from the conduction bands. After you obtained the screened interaction you are ready to build the self-energy operator and solve the corresponding Dyson equation, have a look to the Dyson Equation solvers in Yambo. In order to get the quasi-particle energies, just do yambo -b -g n. After calculation are completed Yambo will produce an output file o.qp which contains the values of the bare and re-normalized energy levels.


I had interpreted "yambo -b -g n" as a calculation with the static Screened EXchange self-energy, without the coulomb hole correlation. Looking at the report file, I see it really does Real Axis GW and it uses the dynamical dielectric matrix db.em1d. After reading the following tutorial (real-axis GW) and running yambo without -b, I realized that it is the same.

All is clear now, until the next opportunity for Murphy's action.

Best wishes
Eduardo
Eduardo Menendez Proupin
Departamento de Química Fisica Aplicada
Facultad de Ciencias
Universidad Autónoma de Madrid
28049 Madrid, Spain
On leave from: Departamento de Fisica, Facultad de Ciencias, Universidad de Chile URL: www.gnm.cl/emenendez
eariel99
 
Posts: 10
Joined: Tue Apr 24, 2012 10:35 pm

Previous

Return to Other issues

Who is online

Users browsing this forum: No registered users and 1 guest

cron